* ARMeilleure: Respect Fz flag for all floating point operations.
This is a change in strategy for emulating the Fz FPCR flag. Before, it was set before instructions that "needed it" and reset after. However, this missed a few hot instructions like the multiplication instruction, and the entirety of A32.
The new strategy is to set the Fz flag only in the following circumstances:
- Set to match FPCR before translated functions/loop are executed.
- Reset when calling SoftFloat methods, set when returning.
- Reset when exiting execution.
This allows us to remove the code around the existing Fz aware instructions, and get the accuracy benefits on all floating point instructions executed while in translated code.
Single step executions now need to be called with a context wrapper - right now it just contains the Fz flag initialization, and won't actually do anything on ARM.
This fixes a bug in Breath of the Wild where some physics interactions could randomly crash the game due to subnormal values not flushing to zero.
This is draft right now because I need to answer the questions:
- Does dotnet avoid changing the value of Mxcsr?
- Is it a good idea to assume that? Or should the flag set/restore be done on every managed method call, not just softfloat?
- If we assume that, do we want a unit test to verify the behaviour?
I recommend testing a bunch of games, especially games affected when this was originally added, such as #1611.
* Remove unused method
* Use FMA for Fmadd, Fmsub, Fnmadd, Fnmsub, Fmla, Fmls
...when available.
Similar implementation to A32
* Use FMA for Frecps, Frsqrts
* Don't set DAZ.
* Add round mode to ARM FP mode
* Fix mistakes
* Add test for FP state when calling managed methods
* Add explanatory comment to test.
* Cleanup
* Add A64 FPCR flags
* Vrintx_S A32 fast path on A64 backend
* Address feedback 1, re-enable DAZ
* Fix FMA instructions By Elem
* Address feedback
* Turn `MemoryOperand` into a struct
* Remove `IntrinsicOperation`
* Remove `PhiNode`
* Remove `Node`
* Turn `Operand` into a struct
* Turn `Operation` into a struct
* Clean up pool management methods
* Add `Arena` allocator
* Move `OperationHelper` to `Operation.Factory`
* Move `OperandHelper` to `Operand.Factory`
* Optimize `Operation` a bit
* Fix `Arena` initialization
* Rename `NativeList<T>` to `ArenaList<T>`
* Reduce `Operand` size from 88 to 56 bytes
* Reduce `Operation` size from 56 to 40 bytes
* Add optimistic interning of Register & Constant operands
* Optimize `RegisterUsage` pass a bit
* Optimize `RemoveUnusedNodes` pass a bit
Iterating in reverse-order allows killing dependency chains in a single
pass.
* Fix PPTC symbols
* Optimize `BasicBlock` a bit
Reduce allocations from `_successor` & `DominanceFrontiers`
* Fix `Operation` resize
* Make `Arena` expandable
Change the arena allocator to be expandable by allocating in pages, with
some of them being pooled. Currently 32 pages are pooled. An LRU removal
mechanism should probably be added to it.
Apparently MHR can allocate bitmaps large enough to exceed the 16MB
limit for the type.
* Move `Arena` & `ArenaList` to `Common`
* Remove `ThreadStaticPool` & co
* Add `PhiOperation`
* Reduce `Operand` size from 56 from 48 bytes
* Add linear-probing to `Operand` intern table
* Optimize `HybridAllocator` a bit
* Add `Allocators` class
* Tune `ArenaAllocator` sizes
* Add page removal mechanism to `ArenaAllocator`
Remove pages which have not been used for more than 5s after each reset.
I am on fence if this would be better using a Gen2 callback object like
the one in System.Buffers.ArrayPool<T>, to trim the pool. Because right
now if a large translation happens, the pages will be freed only after a
reset. This reset may not happen for a while because no new translation
is hit, but the arena base sizes are rather small.
* Fix `OOM` when allocating larger than page size in `ArenaAllocator`
Tweak resizing mechanism for Operand.Uses and Assignemnts.
* Optimize `Optimizer` a bit
* Optimize `Operand.Add<T>/Remove<T>` a bit
* Clean up `PreAllocator`
* Fix phi insertion order
Reduce codegen diffs.
* Fix code alignment
* Use new heuristics for degree of parallelism
* Suppress warnings
* Address gdkchan's feedback
Renamed `GetValue()` to `GetValueUnsafe()` to make it more clear that
`Operand.Value` should usually not be modified directly.
* Add fast path to `ArenaAllocator`
* Assembly for `ArenaAllocator.Allocate(ulong)`:
.L0:
mov rax, [rcx+0x18]
lea r8, [rax+rdx]
cmp r8, [rcx+0x10]
ja short .L2
.L1:
mov rdx, [rcx+8]
add rax, [rdx+8]
mov [rcx+0x18], r8
ret
.L2:
jmp ArenaAllocator.AllocateSlow(UInt64)
A few variable/field had to be changed to ulong so that RyuJIT avoids
emitting zero-extends.
* Implement a new heuristic to free pooled pages.
If an arena is used often, it is more likely that its pages will be
needed, so the pages are kept for longer (e.g: during PPTC rebuild or
burst sof compilations). If is not used often, then it is more likely
that its pages will not be needed (e.g: after PPTC rebuild or bursts
of compilations).
* Address riperiperi's feedback
* Use `EqualityComparer<T>` in `IntrusiveList<T>`
Avoids a potential GC hole in `Equals(T, T)`.
* Optimization | Modify Add Instruction to use LEA instead.
Currently, the add instruction requires 4 registers to take place. By using LEA, we can effectively perform the same working using 3 registers, reducing memory spills and improving translation efficiency.
* Fix IsSameOperandDestSrc1 Check for Add
* Use LEA if Dest != SRC1
* Update IsSameOperandDestSrc1 to account for Cases where Dest and Src1 can be same for add
* Fix error in logic
* Typo
* Add paranthesis for clarity
* Compare registers as requested.
* Cleanup if statement, use same comparison method as generateCopy
* Make change as recommended by gdk
* Perform check only when Add calls are made
* use ensure sametype for lea, fix else
* Update comment
* Update version #
* Add Compare instruction
* Add BranchIf instruction
* Use test when BranchIf & Compare against 0
* Propagate Compare into BranchIfTrue/False use
- Propagate Compare operations into their BranchIfTrue/False use and
turn these into a BranchIf.
- Clean up Comparison enum.
* Replace BranchIfTrue/False with BranchIf
* Use BranchIf in EmitPtPointerLoad
- Using BranchIf early instead of BranchIfTrue/False improves LCQ and
reduces the amount of work needed by the Optimizer.
EmitPtPointerLoader was a/the big producer of BranchIfTrue/False.
- Fix asserts firing when assembling BitwiseAnd because of type
mismatch in EmitStoreExclusive. This is harmless and should not
cause any diffs.
* Increment PPTC interval version
* Improve IRDumper for BranchIf & Compare
* Use BranchIf in EmitNativeCall
* Clean up
* Do not emit test when immediately preceded by and
* Add CRC32 A32 instructions.
* Fix CRC32 instructions.
* Add CRC intrinsic and fast path.
Loop is currently unrolled, will look into adding temp vars after tests are added.
* Begin work on Crc tests
* Fix SSE4.2 path for CRC32C, finialize tests.
* Remove unused IR path.
* Fix spacing between prefix checks.
* This should be Src.
* PTC Version
* OpCodeTable Order
* Integer check improvement. Value and Crc can be either 32 or 64 size.
* This wasn't necessary...
* If size is 3, value type must be I64.
* Fix same src+dest handling for non crc intrinsics.
* Pre-fix (ha) issue with vex encodings
* Start of JIT garbage collection improvements
- thread static pool for Operand, MemoryOperand, Operation
- Operands and Operations are always to be constructed via their static
helper classes, so they can be pooled.
- removing LinkedList from Node for sources/destinations (replaced with
List<>s for now, but probably could do arrays since size is bounded)
- removing params constructors from Node
- LinkedList<> to List<> with Clear() for Operand assignments/uses
- ThreadStaticPool is very simple and basically just exists for the
purpose of our specific translation allocation problem. Right now it
will stay at the worst case allocation count for that thread (so far) -
the pool can never shrink.
- Still some cases of Operand[] that haven't been removed yet. Will need
to evaluate them (eg. is there a reasonable max number of params for
Calls?)
* ConcurrentStack instead of ConcurrentQueue for Rejit
* Optimize some parts of LSRA
- BitMap now operates on 64-bit int rather than 32-bit
- BitMap is now pooled in a ThreadStatic pool (within lrsa)
- BitMap now is now its own iterator. Marginally speeds up iterating
through the bits.
- A few cases where enumerators were generated have been converted to
forms that generate less garbage.
- New data structure for sorting _usePositions in LiveIntervals. Much
faster split, NextUseAfter, initial insertion. Random insertion is
slightly slower.
- That last one is WIP since you need to insert the values backwards. It
would be ideal if it just flipped it for you, uncomplicating things on
the caller side.
* Use a static pool of thread static pools. (yes.)
Prevents each execution thread creating its own lowCq pool and making me cry.
* Move constant value to top, change naming convention.
* Fix iteration of memory operands.
* Increase max thread count.
* Address Feedback
* Implement Jump Table for Native Calls
NOTE: this slows down rejit considerably! Not recommended to be used
without codegen optimisation or AOT.
- Does not work on Linux
- A32 needs an additional commit.
* A32 Support
(WIP)
* Actually write Direct Call pointers to the table
That would help.
* Direct Calls: Rather than returning to the translator, attempt to keep within the native stack frame.
A return to the translator can still happen, but only by exceptionally
bubbling up to it.
Also:
- Always translate lowCq as a function. Faster interop with the direct
jumps, and this will be useful in future if we want to do speculative
translation.
- Tail Call Detection: after the decoding stage, detect if we do a tail
call, and avoid translating into it. Detected if a jump is made to an
address outwith the contiguous sequence of blocks surrounding the entry
point. The goal is to reduce code touched by jit and rejit.
* A32 Support
* Use smaller max function size for lowCq, fix exceptional returns
When a return has an unexpected value and there is no code block
following this one, we now return the value rather than continuing.
* CompareAndSwap (buggy)
* Ensure CompareAndSwap does not get optimized away.
* Use CompareAndSwap to make the dynamic table thread safe.
* Tail call for linux, throw on too many arguments.
* Combine CompareAndSwap 128 and 32/64.
They emit different IR instructions since their PreAllocator behaviour
is different, but now they just have one function on EmitterContext.
* Fix issues separating from optimisations.
* Use a stub to find and execute missing functions.
This allows us to skip doing many runtime comparisons and branches, and reduces the amount of code we need to emit significantly.
For the indirect call table, this stub also does the work of moving in the highCq address to the table when one is found.
* Make Jump Tables and Jit Cache dynmically resize
Reserve virtual memory, commit as needed.
* Move TailCallRemover to its own class.
* Multithreaded Translation (based on heuristic)
A poor one, at that. Need to get core count for a better one, which
means a lot of OS specific garbage.
* Better priority management for background threads.
* Bound core limit a bit more
Past a certain point the load is not paralellizable and starts stealing from the main thread. Likely due to GC, memory, heap allocation thread contention. Reduce by one core til optimisations come to improve the situation.
* Fix memory management on linux.
* Temporary solution to some sync problems.
This will make sure threads exit correctly, most of the time. There is a potential race where setting the sync counter to 0 does nothing (counter stays at what it was before, thread could take too long to exit), but we need to find a better way to do this anyways. Synchronization frequency has been tightened as we never enter blockwise segments of code. Essentially this means, check every x functions or loop iterations, before lowcq blocks existed and were worth just as much. Ideally it should be done in a better way, since functions can be anywhere from 1 to 5000 instructions. (maybe based on host timer, or an interrupt flag from a scheduler thread)
* Address feedback minus CompareAndSwap change.
* Use default ReservedRegion granularity.
* Merge CompareAndSwap with its V128 variant.
* We already got the source, no need to do it again.
* Make sure all background translation threads exit.
* Fix CompareAndSwap128
Detection criteria was a bit scuffed.
* Address Comments.
* Start of the ARMeilleure project
* Refactoring around the old IRAdapter, now renamed to PreAllocator
* Optimize the LowestBitSet method
* Add CLZ support and fix CLS implementation
* Add missing Equals and GetHashCode overrides on some structs, misc small tweaks
* Implement the ByteSwap IR instruction, and some refactoring on the assembler
* Implement the DivideUI IR instruction and fix 64-bits IDIV
* Correct constant operand type on CSINC
* Move division instructions implementation to InstEmitDiv
* Fix destination type for the ConditionalSelect IR instruction
* Implement UMULH and SMULH, with new IR instructions
* Fix some issues with shift instructions
* Fix constant types for BFM instructions
* Fix up new tests using the new V128 struct
* Update tests
* Move DIV tests to a separate file
* Add support for calls, and some instructions that depends on them
* Start adding support for SIMD & FP types, along with some of the related ARM instructions
* Fix some typos and the divide instruction with FP operands
* Fix wrong method call on Clz_V
* Implement ARM FP & SIMD move instructions, Saddlv_V, and misc. fixes
* Implement SIMD logical instructions and more misc. fixes
* Fix PSRAD x86 instruction encoding, TRN, UABD and UABDL implementations
* Implement float conversion instruction, merge in LDj3SNuD fixes, and some other misc. fixes
* Implement SIMD shift instruction and fix Dup_V
* Add SCVTF and UCVTF (vector, fixed-point) variants to the opcode table
* Fix check with tolerance on tester
* Implement FP & SIMD comparison instructions, and some fixes
* Update FCVT (Scalar) encoding on the table to support the Half-float variants
* Support passing V128 structs, some cleanup on the register allocator, merge LDj3SNuD fixes
* Use old memory access methods, made a start on SIMD memory insts support, some fixes
* Fix float constant passed to functions, save and restore non-volatile XMM registers, other fixes
* Fix arguments count with struct return values, other fixes
* More instructions
* Misc. fixes and integrate LDj3SNuD fixes
* Update tests
* Add a faster linear scan allocator, unwinding support on windows, and other changes
* Update Ryujinx.HLE
* Update Ryujinx.Graphics
* Fix V128 return pointer passing, RCX is clobbered
* Update Ryujinx.Tests
* Update ITimeZoneService
* Stop using GetFunctionPointer as that can't be called from native code, misc. fixes and tweaks
* Use generic GetFunctionPointerForDelegate method and other tweaks
* Some refactoring on the code generator, assert on invalid operations and use a separate enum for intrinsics
* Remove some unused code on the assembler
* Fix REX.W prefix regression on float conversion instructions, add some sort of profiler
* Add hardware capability detection
* Fix regression on Sha1h and revert Fcm** changes
* Add SSE2-only paths on vector extract and insert, some refactoring on the pre-allocator
* Fix silly mistake introduced on last commit on CpuId
* Generate inline stack probes when the stack allocation is too large
* Initial support for the System-V ABI
* Support multiple destination operands
* Fix SSE2 VectorInsert8 path, and other fixes
* Change placement of XMM callee save and restore code to match other compilers
* Rename Dest to Destination and Inst to Instruction
* Fix a regression related to calls and the V128 type
* Add an extra space on comments to match code style
* Some refactoring
* Fix vector insert FP32 SSE2 path
* Port over the ARM32 instructions
* Avoid memory protection races on JIT Cache
* Another fix on VectorInsert FP32 (thanks to LDj3SNuD
* Float operands don't need to use the same register when VEX is supported
* Add a new register allocator, higher quality code for hot code (tier up), and other tweaks
* Some nits, small improvements on the pre allocator
* CpuThreadState is gone
* Allow changing CPU emulators with a config entry
* Add runtime identifiers on the ARMeilleure project
* Allow switching between CPUs through a config entry (pt. 2)
* Change win10-x64 to win-x64 on projects
* Update the Ryujinx project to use ARMeilleure
* Ensure that the selected register is valid on the hybrid allocator
* Allow exiting on returns to 0 (should fix test regression)
* Remove register assignments for most used variables on the hybrid allocator
* Do not use fixed registers as spill temp
* Add missing namespace and remove unneeded using
* Address PR feedback
* Fix types, etc
* Enable AssumeStrictAbiCompliance by default
* Ensure that Spill and Fill don't load or store any more than necessary
